Control Profiles of Complex Networks

Justin Ruths and Derek Ruths

Science 343, 1373 (2014);
DOI: 10.1126/science.1242063

This copy is for your personal, non-commercial use only.

If you wish to distribute this article to others, you can order high-quality copies for your
colleagues, clients, or customers by clicking here.
Permission to republish or repurpose articles or portions of articles can be obtained by
following the guidelines here.

The following resources related to this article are available online at

www.sciencemag.org (this information is current as of April 2, 2014 ):

Downloaded from www.sciencemag.org on April 2, 2014

Updated information and services, including high-resolution figures, can be found in the online
version of this article at:
http://www.sciencemag.org/content/343/6177/1373.full.html
Supporting Online Material can be found at:
http://www.sciencemag.org/content/suppl/2014/03/19/343.6177.1373.DC1.html
A list of selected additional articles on the Science Web sites related to this article can be
found at:
http://www.sciencemag.org/content/343/6177/1373.full.html#related
This article cites 47 articles, 9 of which can be accessed free:
http://www.sciencemag.org/content/343/6177/1373.full.html#ref-list-1
This article has been cited by 1 articles hosted by HighWire Press; see:
http://www.sciencemag.org/content/343/6177/1373.full.html#related-urls
This article appears in the following subject collections:
Sociology
http://www.sciencemag.org/cgi/collection/sociology

Science (print ISSN 0036-8075; online ISSN 1095-9203) is published weekly, except the last week in December, by the
American Association for the Advancement of Science, 1200 New York Avenue NW, Washington, DC 20005. Copyright
2014 by the American Association for the Advancement of Science; all rights reserved. The title Science is a
registered trademark of AAAS.
 REPORTS
text). Some of these dilations may arise due to a
Control Profiles of Complex Networks surplus of sink nodes (Nt > Ns). In this case
we call the dilations due to surplus sink nodes,
Justin Ruths1* and Derek Ruths2 Ne = max(0, Nt – Ns), external dilation points (see
the branching at node E to F and G in Fig. 1). The
Studying the control properties of complex networks provides insight into how designers and remaining stem dilations give rise to a class of
engineers can influence these systems to achieve a desired behavior. Topology of a network has controls due to more nuanced network structures,
been shown to strongly correlate with certain control properties; here we uncover the fundamental called internal dilation points, Ni (see the branch-
structures that explain the basis of this correlation. We develop the control profile, a statistic ing at node A to nodes B and C in Fig. 1).
that quantifies the different proportions of control-inducing structures present in a network. We Taken together, the minimum number of in-
find that standard random network models do not reproduce the kinds of control profiles that dependent controls (Nc) required for full control
are observed in real-world networks. The profiles of real networks form three well-defined clusters of a complex network is, then, the sum of the
that provide insight into the high-level organization and function of complex systems. number of source nodes, external dilation points,
and internal dilation points,
any natural complex systems are im- paths, called stems, each driven by an indepen-

M portant or vital to human life, society,

and governance. Undesirable behavior
of such systems, observed in the form of disease,
dent control (6). These paths can then control cycles
that are inherently self-regulatory (see supplemen-
tary text). However, ultimately it is these stems that
Nc = Ns + Ne + Ni

with Ns, Ne, and Ni not all zero (4, 9). One way to
(1)

epidemics, economic collapse, and social unrest, dictate the need for controls: There must be one calculate Nc directly is by using a maximum match-
has generated considerable interest in being able control for each stem in the system (7). ing algorithm, which, while accurate, is expensive
to control complex systems modeled as networks. An immediate question, then, is how many for large networks (see supplementary text).
Previous work has already established the con- stems exist in the system. The location of stems is With this understanding, we can examine the
nection between control theory and complex largely dictated by two topological network fea- breakdown of the origin of controls in terms of
networks (1–3). Controlling a complex directed tures: sources, which are nodes with edges only amounts of source nodes, external dilation points,
network reduces to (i) the selection of specific pointing away from them (e.g., node A in Fig. 1), and internal dilation points for different classes
nodes that, by virtue of being directly controlled, and sinks, which are nodes with edges only point- of real-world networks, as well as for widely used
can indirectly drive the rest of the network to any ing toward them (e.g., nodes F and G in Fig. 1). A generative network models (see supplementary text
arbitrary state and (ii) the prescription of time- source, because it does not have edges entering it, for the complete listing). To make direct compar-
varying control inputs that, when applied to these can only appear at the origin of a stem and there- isons across networks of different size, in the fol-
nodes, will drive the system to a desired state. fore must be directly controlled. Similarly, a sink, lowing discussion we normalize by the number
Work to date has focused on the number of lacking edges to influence other nodes, must lie of nodes, N, i.e., (ns, ne, ni) = (Ns /N, Ne /N, Ni /N)
controls and their attachment points required by a at the end of a stem. The number of stems in a and nc = Nc /N.
complex network (4). Existing studies have shown complex system, then, is at least max(Ns, Nt), where For any network, ncse = ns + ne can be com-
how the mathematics of structural controllability Ns and Nt are the number of sources and sinks, re- puted directly from the network’s degree dis-
can be used to identify a minimum number of spectively. Without lack of generality, we consider tribution, namely, the sum of the fraction of nodes
controls that can drive the system as a whole to here a graph with no disconnected nodes (8). with in-degree zero and out-degree zero. This ex-
any arbitrary state. A main outcome of past work In situations where a path must branch into plains why Liu et al. found that the degree dis-
revealed that the degree distribution of a network two or more paths in order to reach all nodes, a tribution of a network correlates so strongly with
alone (i.e., the probability distribution over the dilation occurs: Additional stems must be intro- the number of controls (1). Our findings show
number of links that nodes have) is correlated duced at these branching points, each one re- that knowing the full degree distribution is often
with the minimum number of controls (1). quiring a separate control (see supplementary unnecessary since, on average, the number of con-
Understanding the control properties of a com-
plex system requires not just knowing how many
controls are needed, but also characterizing the
functional origin of each control, an explanatory
detail not provided by the degree distribution cor-
relation. For example, a biochemical system and a
financial system might require the same number
of controls, but the structures within the networks
that give rise to these controls could be quite differ-
ent. Ultimately, planning and implementing a control
scheme on a network will depend greatly on how
these control points are situated in the network.
The formulation of control that we employ,
structural controllability, formalizes the idea that,
in the absence of a feedback loop (or, more gen-
erally, a cycle), a given node can control at most
one of its immediate neighbors (5). Thus, the struc- Fig. 1. An example of (left) a simple directed network and (right) one of the possible minimum
ture of the propagation of control influence through control structures, constructed of stems (blue) and cycles (purple and green). There is one source
the network consists of a backbone of directed (node A) and two sinks (nodes F and G). Three controls are required arising from one source (node A), one
external dilation (due to two sink nodes F and G), and one internal dilation (due to the A-B-C-D topology).
1
Engineering Systems and Design, Singapore University of The dashed lines indicate how control signals enter the network. For control 1, the same signal is sent to
Technology and Design, Singapore. 2Department of Computer nodes A and K. Controls 2 and 3 send different signals to nodes B and F, respectively. Gray arrows
Science, McGill University, Montreal, Quebec, Canada. indicate edges that are not involved in the control structure of the network. See the supplementary text for
*Corresponding author. E-mail: justinruths@sutd.edu.sg additional details regarding the determination of the control structure and its various components.

www.sciencemag.org SCIENCE VOL 343 21 MARCH 2014 1373

 REPORTS
trols is dominated by only two points in the degree computational steps proportional to the number cause each of these terms can be directly com-
distribution: sources and sinks (i.e., nodes with of edges, which can be quadratically greater than puted. We obtain the control profile for a complex
in-degree = 0 or out-degree = 0, respectively). the number of nodes (see supplementary text). network, given by the triple (hs, he, hi). Using these
Indeed, the fraction of nodes that are internal Moreover, this estimate performs as well as one control profiles and their corresponding plots (see
dilation points (ni = nc – ncse) is, in many cases, based on the full degree distribution, despite re- fig. S4), we can investigate the extent to which
relatively small in both real networks and the ran- quiring much less information. In Fig. 2, we com- the profiles of networks differ.
dom network models that are often used to ap- pare the predictive power of using ns and ne alone We have already established that, of the syn-
proximate real-world networks. Across a total of (ncse, blue) against the number of controls implied thetic networks considered, few have appreciable
70 real networks—sampled from ecological, eco- by the network’s full degree distribution (ncdeg, red). numbers of internal or external dilations. This is
nomic, neural, organizational, social, transcriptional, We find that, on average across all the surveyed largely due to the way in which these networks
and technological systems (see table S2)—ni av- networks, the additional information from the full are generated. The generative models that yielded
eraged 14.0%. We observed low ni across ran- degree distribution accounts for only 5.1% im- networks with few internal dilations had one of
dom networks as well: 3.0, 3.0, 2.8, and 24.9% in provement in the prediction for real networks two key features. BA and LA networks grow by
Erdos-Renyi (ER), Barabasi-Albert (BA), local and 2.7% (ER), 0.6% (BA), 0.7% (LA), and 7.3% new node attachment: At each time step, a new
attachment (LA), and duplication-divergence (DD) (DD) for synthetic networks (see fig. S1 for further node is added that has no inbound neighbors.
networks, respectively (10–13). When ni is small, details). This further indicates that additional sta- Preferential attachment further does not favor
as is the case in many networks, the bound given tistics beyond the degree distribution will be needed creating links to these new sources, yielding an
by ncse is quite tight. Notably, DD networks ex- to fully capture the number of internal dilations enrichment in sources. ER, by contrast, creates
hibit a higher fraction of internal dilation points and to reduce these prediction errors. sources and sinks with equal probability, making
(ni = 24.9%), but still small enough to meaning- Although prediction is a powerful use for our it difficult for generated networks to have sub-
fully employ the estimates above. framework, we now return to the original ques- stantial numbers of external dilations.
By simply counting source and sink nodes tion regarding the decomposition of controls in By definition, a dilation requires an expansion
(a task that is linear in the number of nodes), networks. To make an equitable comparison across point in the network—a region where a smaller
we obtain a relatively good lower bound on the networks with different numbers of controls, but number of nodes link into a larger number of nodes.
number of controls. This approach is a marked potentially similar relative composition of con- Preferential attachment is biased against expansion,
improvement over the maximum matching al- trols, we modify Eq. 1 to represent the fractions of favoring linking into smaller populations of high-
gorithm, which is used to obtain the exact num- controls due to source nodes (hs = Ns/Nc), external degree nodes. ER also has no built-in bias toward
ber of minimum controls. This algorithm requires dilation points (he = Ne/Nc), and internal dilation expansion points. DD, however, does, by virtue of
points (hi = Ni/Nc) by normalizing the equation creating new nodes through copy events (14). Over-
with respect to the number of controls, Nc: all, this suggests that the expected values of a control
1
profile may be anticipated and understood by ap-
hs + he + hi = 1 (2) pealing to the formation mechanism that gen-
erated it rather than particular high-level network
Because Nc can be computed by means of a properties that the mechanism induces.
|nc - ncse | |nc - ncdeg |

maximum matching algorithm (see supplementary Real networks have control profiles that are
text), and Ns and Ne are obtained from analysis vastly different from the synthetic networks con-
of node degrees, we can trivially solve for Ni. Our sidered (see Fig. 3). This is related to the obser-
goal is no longer to estimate Nc; hence, defining Ni vation above—that existing generation methods
as a function of Ns, Ne, and Nc is not circular be- are biased toward source-based controls. How-

source external-dilation internal-dilation

dominated dominated dominated

0
ER BA LA DD real
Fig. 2. Source and sink nodes dominate the
correlation between degree distribution and the
number of required controls. The error in the ap-
real
proximation of the fraction of required controls is only
slightly improved by incorporating the nonzero terms neural (4) social (8) copurchase (4) corp. ownership (1) airport (2) autonomous (1)
of the degree distribution (ncdeg = Ncdeg/N is the
fraction of required controls after a degree-preserving
shuffle; ncse = ns + ne is the fraction of source and
sink nodes in the network). The mean improvement ER BA messaging (2) p2p (9) circuit (3) food web (22)
in approximation error for synthetic networks—Erdos-
Renyi (ER), Barabasi-Albert (BA), local attachment (LA),
and duplication-divergence (DD)—and real networks
is indicated by the difference in bar heights. Synthetic
synthetic LA DD social influence (6) transcription (2) www+blog (4)
networks were generated from a comprehensive set of
parameters with network sizes, N, ranging from 100 to Fig. 3. Real and synthetic networks have substantially different control profiles. The coloring of
50,000; average degree, L/N, ranging from 2 to 58; and breakout subfigures indicates the clustering observed in various types of real networks. Deeper shades of
for DD networks, a probability to retain copied edges the heatmap indicate a greater density of networks with control profiles located in that region. Networks
from 0.1 to 0.9 (see supplementary text). The error bars with Ns = Nt = Ni = 0 are omitted from this figure (9). The aggregated heatmaps are normalized to avoid
represent one standard deviation from the mean ap- biases induced by over- and underrepresented types of networks (see supplementary text for details).
proximation error. Numbers in parentheses indicate numbers of networks present in each plot.

1374 21 MARCH 2014 VOL 343 SCIENCE www.sciencemag.org

 REPORTS
ever, more generally, these methods lack the abil- are structured to enable independent computa- trol properties. Certainly, the scale and organic
ity to specify a target control profile at all. Thus, tions and interaction at a massive scale (20, 21). nature of connectivity present in many complex
in much the same way as many network genera- Internal-dilation–dominated networks include systems suggest that strategies for controlling
tion methods have parameters to set specific de- food networks, electrical circuits, airport inter- them will be nontrivial; thus, such fundamental
gree distributions or clustering, future work on connectivity, intranets, and the Internet’s autono- insights may be crucial to such endeavors.
controllability in complex networks will require mous systems. These share a fundamental feature in Here, we have identified the major structures
network models in which the control profile can common: They are all, more or less, closed systems. that induce the need for controls in a complex
be tuned to match those observed in the real world. Reinforcing this idea is that many of these systems network in order to guarantee controllability. This
Real networks also have control profiles that obey conservation laws, i.e., Kirchhoff’s law for elec- led to the discovery that for many networks, we
are different from the profiles of their random- trical circuits and biomass conservation within an can approximate the number of controls simply
ized counterparts. Figure S3 shows the impact ecosystem. Those that do not have a formal conser- by the zero-degree points of the degree distribu-
of several randomization schemes, in particular, vation law are still characterized by the circulation tion. Subsequently, we defined the control pro-
highlighting the loss or gain of internal dilations and recirculation of a resource (e.g., information and file, the decomposition of a complex network
indicated visually by a vertical shift in the heatmap perspectives in the blogosphere and population into the structural elements that induce the need
density. This suggests that many real networks movement within transportation systems) that drive for specific controls—sources, external dilations,
have an internal dilation signature that distinguishes the evolution of the system’s state. Consistent with and internal dilations. We find that control pro-
them from other randomized instantiations. the notion of a closed system and conservation files highlight a profound and systemic difference
Although the control profiles of real networks are laws, internal-dilation–dominated networks rep- between the control structures of random net-
diverse, they also show a strong tendency to cluster resent systems that can be intended to function work models and real networks. Furthermore, the
around the three components of the control profile relatively independently of external stimulation. control profiles of real networks fall into three
itself, suggesting that real networks broadly fall into More generally, external-dilation– and internal- clusters that correspond to different aspects of
three distinct classes: external-dilation dominated, dilation–dominated networks both require many high-level structure and function. Thus, the con-
source dominated, and internal-dilation dominated controls to be attached to nonsource nodes in order trol profile is a network statistic that may prove
(see Fig. 3). Because these clusters do not corre- to provide full controllability. When sources corre- useful in better understanding and approaching
spond to known groupings of networks, we can spond to inputs to the system, we might interpret the control of complex networks.
test the conjecture that similar network types cluster dilations to be the signature of a network that is not
together. These control profile clusters are statisti- meant to be fully controlled, in that it is not de- References and Notes
cally significant according to two separate tests for signed to visit every possible system state (e.g., the 1. Y.-Y. Liu, J.-J. Slotine, A.-L. Barabási, Nature 473,
such consistency (see table S1 and fig. S2). conservation laws in internal-dilation–dominated 167–173 (2011).
2. M. Pósfai, Y.-Y. Liu, J.-J. Slotine, A.-L. Barabási, Sci. Rep.
Although the class of external-dilation– networks impose constraints such that only a por- 3, 1067 (2013).
dominated networks comprises a diverse set of tion of the full state space can be explored). Indeed, 3. T. Jia et al., Nat. Commun. 4, 2002 (2013).
systems, the control profiles of these networks in the case of transcriptional networks, electronic 4. In this work, we discuss the minimum number of
consistently reflect a surplus of sinks, indicating systems, and the Internet, certain states are path- independent controls (denoted Nc). In some previous
literature, this number was called the minimum number
that the systems cannot be fully controlled in a ological to or are simply unexplored by a system of driver nodes (denoted ND) (1).
top-down manner (from sources alone). Instead, under natural conditions (22). A tantalizing direc- 5. C.-T. Lin, IEEE Trans. Automat. Contr. AC-19, 201–208
a control scheme that uses only sources will yield tion for future work is to consider the extent to (1974).
6. The structure capturing the propagation of control,
correlated behavior among nodes that are down- which internal dilations themselves reveal char- composed of stems, cycles, and buds, is called cacti. Each
stream from a common source. Such synchronized acteristics of forbidden regions of state space. cactus within the cacti contains only one stem and any
behavior may be an objective of the system. For Nevertheless, the extent to which the control number of cycles and buds (see supplementary text).
example, in corporate-ownership, leadership, and properties of networks are determined by source 7. C. Commault, J.-M. Dion, J. W. Van der Woude,
Kybernetika 38, 503–520 (2002).
trust hierarchies, decisions or beliefs at the sources nodes can give insight into whether it is actually 8. An immediate extension can be made to accommodate
(the owners and leaders) will drive correlated be- practical to control the entire system. Sources in a disconnected nodes as stems of unit length, acting as
havior, functional differentiation, and beliefs in network can correspond to elements that lie on the both a source and sink.
owned organizations, followers, and advice-seekers boundary of the system—variables to which we 9. In the exceptional case where Ns = Ne = Ni = 0, at
least one control must be present, so Nc = 1. Formally
(15, 16). Genes involved in transcriptional sys- have direct access. For example, in the case of bio- we could revise Eq. 1 to Nc = max(1, Ns + Ne + Ni).
tems exhibit such high degrees of correlated ex- chemical pathways in human cells, target receptor See supplementary text for more discussion.
pression that these correlations form the basis for proteins responsible for transducing extracellular 10. P. Erdös, A. Renyi, Publicationes Mathematicae 6,
cancer diagnosis and staging (17–19). Similarly, stimuli into intracellular signals act as source nodes 290–297 (1959).
11. A.-L. Barabási, R. Albert, Science 286, 509–512 (1999).
in copurchase networks, many products (accessories) in the cellular signaling network (23, 24). Because 12. M. O. Jackson, B. W. Rogers, Am. Econ. Rev. 97, 890–915
may derive meaning, and therefore shopper rele- these nodes are readily accessible and influence (2007).
vance, from a smaller set of core products (the items the protein interactions within the cell, a large per- 13. I. Ispolatov, P. L. Krapivsky, A. Yuryev, Phys. Rev. E Stat.
Nonlin. Soft Matter Phys. 71, 061911 (2005).
that need to be accessorized). As a result, in each centage of drugs used today target these receptor 14. Edge orientation is a matter of convention; see
case above, to achieve full controllability, additional proteins, rather than intracellular proteins directly, supplementary text for discussion on the sensitivity of
controls must be added to nonsource nodes. to effect change within the cell (25). Thus, networks our results to edge reversals.
Source-dominated networks, unlike the external- whose control profiles are dominated by sources 15. S. Vitali, J. B. Glattfelder, S. Battiston, PLOS ONE 6,
e25995 (2011).
dilation–dominated class, have no more sinks than may have a much more feasible actuation scheme. 16. N. E. Friedkin, A Structural Theory of Social Influence
controls. Where an external-dilation–dominated Finally, despite surveying a wide array of com- (Cambridge Univ. Press, Cambridge, 2006).
control profile indicated a top-down organization, plex networks, our analysis yields few networks 17. N. Bhardwaj, K.-K. Yan, M. B. Gerstein, Proc. Natl. Acad.
intended to drive correlated behavior through the with profiles that contain nontrivial contributions of Sci. U.S.A. 107, 6841–6846 (2010).
18. H. Yu, M. Gerstein, Proc. Natl. Acad. Sci. U.S.A. 103,
system, a source-dominated control profile may all three components. This suggests that many com- 14724–14731 (2006).
suggest the opposite—a tendency toward allowing plex systems may share in common a general prin- 19. L. J. van ’t Veer et al., Nature 415, 530–536 (2002).
relatively uncorrelated behavior. This makes dis- ciple that makes such control profiles unfavorable. 20. E. Bullmore, O. Sporns, Nat. Rev. Neurosci. 10, 186–198
(2009).
tributed processing possible, which is characteristic Understanding the causal origin of a network’s 21. M. Girvan, M. E. J. Newman, Proc. Natl. Acad. Sci. U.S.A.
of the systems that fall into this class: biological minimal control points can yield both direct and 99, 7821–7826 (2002).
neural and social networks. In both, the systems comparative insights into the system and its con- 22. F.-J. Müller, A. Schuppert, Nature 478, E4 (2011).

www.sciencemag.org SCIENCE VOL 343 21 MARCH 2014 1375

 REPORTS
23. W. K. Kroeze, D. J. Sheffler, B. L. Roth, J. Cell Sci. 116, this manuscript. Supported by the Singapore University of Technology Figs. S1 to S4
4867–4869 (2003). and Design–Massachusetts Institute of Technology International Tables S1 and S2
24. J. S. Gutkind, Sci. STKE 2000, re1 (2000). Design Center (IDG31300103) and by Natural Sciences and References (26–70)
25. M. J. Marinissen, J. S. Gutkind, Trends Pharmacol. Sci. Engineering Research Council (Discovery Grant 125517855).
22, 368–376 (2001).
Supplementary Materials
Acknowledgments: We thank the anonymous reviewers for their www.sciencemag.org/content/343/6177/1373/suppl/DC1 18 June 2013; accepted 31 January 2014
thoughtful and insightful critiques, which substantively improved Materials and Methods 10.1126/science.1242063

Fossilized Nuclei and Chromosomes of geologic time (8). To date, evidence for evo-
lutionary conservatism in fern genomes has been
exclusively based on studies of extant plants
Reveal 180 Million Years of (9, 10). Here, we present direct paleontological
evidence for long-term genomic stasis in this
Genomic Stasis in Royal Ferns family in the form of a calcified osmundaceous
rhizome from the Lower Jurassic of Sweden with
pristinely preserved cellular contents, including
Benjamin Bomfleur,1* Stephen McLoughlin,1* Vivi Vajda2 nuclei and chromosomes.
The specimen was collected from mafic vol-
Rapidly permineralized fossils can provide exceptional insights into the evolution of life over geological caniclastic rocks [informally named the “Djupadal
time. Here, we present an exquisitely preserved, calcified stem of a royal fern (Osmundaceae) formation” (11)] at Korsaröd near Höör, Scania,
from Early Jurassic lahar deposits of Sweden in which authigenic mineral precipitation from Sweden [fig. S1 of (12)]. Palynological analysis in-
hydrothermal brines occurred so rapidly that it preserved cytoplasm, cytosol granules, nuclei, and even dicates an Early Jurassic (Pliensbachian) age for
chromosomes in various stages of cell division. Morphometric parameters of interphase nuclei match these deposits (11) (fig. S2), which agrees with
those of extant Osmundaceae, indicating that the genome size of these reputed “living fossils” has radiometric dates obtained from nearby volcanic
remained unchanged over at least 180 million years—a paramount example of evolutionary stasis. necks (13) from which the basaltic debris originated.
The fern rhizome was permineralized in vivo by
oyal ferns (Osmundaceae) are a basal mundaceous rhizomes from the Mesozoic are calcite from hydrothermal brines (11, 14) that per-

R group of leptosporangiate ferns that have

undergone little morphological and an-
atomical change since Mesozoic times (1–6).
practically indistinguishable from those of mod-
ern genera (3–5) or species (6). Furthermore, with
the exception of one natural polyploid hybrid
1
Department of Palaeobiology, Swedish Museum of Natural
History, Post Office Box 50007, SE-104 05 Stockholm, Sweden.
Well-preserved fossil plants from 220-million- (7), all extant Osmundaceae have an invariant 2
Department of Geology, Lund University, Sölvegatan 12,
year-old rocks already exhibit the distinctive ar- and unusually low chromosome count (7, 8), sug- SE-223 62 Lund, Sweden.
chitecture of the extant interrupted fern (Osmunda gesting that the genome structure of these ferns *Corresponding author. E-mail: benjamin.bomfleur@
claytoniana) (2), and many permineralized os- may have remained unchanged over long periods nrm.se (B.B.); steve.mcloughlin@nrm.se (S.M.)

Fig. 1. Cytological features preserved in the apical region

of the Korsaröd fern fossil. (A) transverse section through
the rhizome; (B) detail of radial longitudinal section showing
typical pith-parenchyma cells with preserved cell membranes
(arrow), cytoplasm and cytosol particles, and interphase nuclei
with prominent nucleoli; (C) interphase nucleus with nucleolus
and intact nuclear membrane; (D) early prophase nucleus with
condensing chromatin and disintegrating nucleolus and
nuclear membrane; (E and F) late prophase cells with coiled
chromosomes and with nucleolus and nuclear membrane
completely disintegrated; (G and H) prometaphase cells
showing chromosomes aligning at the equator of the nucleus;
(I and J) possible anaphase cells showing chromosomes at-
tenuated toward opposite poles. (A), (C to E), (G), and (I)
are from NRM S069656. (B), (F), (H), and (J) are from NRM
S069658. Scale bars: (A) 500 mm; (B) 20 mm; (C to J) 5 mm.

1376 21 MARCH 2014 VOL 343 SCIENCE www.sciencemag.org